Near-infrared fusion for deep lightness enhancement

Lightness enhancement is a long-standing research topic in computer vision. Existing deep learning-based approaches usually extract features from the low-light image to model the enlightening process, which may fall short of robustness since low-light features can be unreliable in heavily dark regions. Inspired by the fact that infrared imaging is immune to illumination variation, we propose to exploit an extra infrared image to help brighten the low-light one. Specifically, we design a deep convolutional neural network to jointly extract the infrared and low-light features and produce a normal-light image under the supervision of multi-scale loss functions, including a discriminator loss that enforces the network output image to mimic a real one. Moreover, a contextual attention module is proposed to reconstruct reliable low-light features in heavily dark regions by exploring feature correlation consistency among low-light and infrared features. Extensive experiments on two composited and one real-world datasets demonstrate the superiority of the proposed approach over existing methods qualitatively and quantitatively.